Automatic Left Ventricular Segmentation In Echocardiography Based On YOLOv3 Model For Constraint And Localization

Cardiovascular disease is the most common and deadly disease,and with the increase in the proportion of aging population in China,the incidence of cardiovascular disease is also increasing year by year,and early diagnosis is of utmost importance for disease prevention.Based on the completion of accurate segmentation of the myocardial wall of the left ventricle,it is important to obtain indicators such as myocardial tissue strain and strain rate for the assessment and analysis of cardiac function.The manual segmentation method is mostly used in clinical practice,which is time-consuming and tedious,and the segmentation results are easily influenced by subjective factors.Therefore,automatic segmentation of echocardiograms has been the focus of research in this field.To address the problems of reliability as well as low accuracy of low-contrast and high-noise echocardiographic image segmentation,we propose to realize the localization of the three locations of the apical and bottom of the left ventricle and the left ventricular region based on the YOLO V3 model and generate the left ventricular subimages,respectively;on this basis,the preliminary identification and binarization of the myocardium of the left ventricular subimages based on the Markov random field model;in the three locations of the left ventricular Under the constraint limitation of the three locations of the left ventricle,we combine nonlinear least squares curve fitting and edge approximation to achieve the accurate segmentation and extraction of the left ventricular endocardium.The specific research includes the following aspects.Firstly,The YOLO V3 model-based implementation was designed to localize the three locations of the left ventricular apex and base as well as the left ventricular region.To address the problem that the left ventricular position cannot be accurately identified due to the large differences in the shape and appearance of the left ventricle in different individuals,the YOLO V3 model is introduced to extract the features of the left ventricular region,to realize the localization of the three positions of the apical and bottom of the left ventricle and the left ventricular region,and to generate left ventricular subimages.The experimental results show that the YOLO V3 model has a strong ability to extract features of the left ventricular region and a limited constraint on the left ventricle,with AP50 values above 92% for all four target regions.Secondly,an algorithm for binarization of echocardiographic left ventricular subimages based on a Markov random field model is proposed.To reduce the effects of echocardiographic scatter and noise,the generated left ventricular subimages were first smoothed and noise-reduced by a two-dimensional Wiener filtering adaptive filtering method;then the initial identification and binarization of the myocardium of the left ventricular subimages were performed by applying a Markov random field-based model,which was optimized using an iterative conditional mode algorithm to make the energy function obtain a minimum value.The experimental results show that the algorithm has certain robustness and better binarization accuracy,and its DICE index value reaches 0.88 ±0.03.Thirdly,A combination of position-constrained frames and nonlinear least squares were used to generate new left ventricular subimages for curve fitting of the myocardial region.In this part,the YOLO V3 model is used to generate three positional constraint boxes under the limiting constraint,combined with nonlinear least squares to generate new left ventricular subimages for curve fitting,locating and limiting the myocardial region in the left ventricle.Finally,the accurate segmentation and extraction of the left ventricular endocardium is achieved based on the fitted curve,the proposed edge approximation algorithm,and the B-Spline algorithm.The experimental results show that the algorithm in this paper has certain robustness and better segmentation accuracy,and the evaluation metrics Computation Speed(fps),Dice,MAD,and HD reach 2.1～2.25,93.578±1.97,2.57±0.89,and 6.68±1.78,respectively.This project addresses the problems of inaccurate echocardiographic left ventricular localization and edge artifact interference,and proposes an automatic echocardiographic left ventricular segmentation based on the YOLO V3 model for constraint and localization.Several sets of experiments have shown that the echocardiographic segmentation algorithm proposed in this paper has good segmentation accuracy and robustness,and the method is particularly advantageous in terms of computational speed,which is very important for the real-time performance required to assess cardiac function based on echocardiography,and the method can achieve better segmentation results using less training data,and lay a better foundation for subsequent auxiliary diagnostic work.